Forecasting step-like landslide displacement through diverse monitoring frequencies

Abstract

The accuracy of displacement prediction plays a pivotal role in landslide prevention and control efforts. However, many existing studies have overlooked the influence of surface monitoring frequency on displacement prediction accuracy. In this study, we investigate the impact of different monitoring frequencies on displacement prediction accuracy using the Baijiabao landslide in the Three Gorges Reservoir Area (TGRA) as a case study. We gathered landslide surface automatic monitoring data at varying monitoring frequencies, including daily, seven days, nine days, eleven days, thirteen days, fifteen days, twenty-one days, and thirty days. To analyze the data, we employed the Ensemble Empirical Mode Decomposition (EEMD) algorithm to decompose accumulated displacements into periodic term displacements and trend term displacements at each monitoring frequency. Subsequently, we predicted the trend term displacement using polynomial fitting, while the periodic term displacement was forecasted using two neural network models: the Long Short-Term Memory model (LSTM) and the Gated Recurrent Unit model (GRU). These predictions were then combined to obtain cumulative displacement predictions, allowing us to compare the prediction accuracies across different monitoring frequencies. Our findings indicate that the proposed prediction models exhibit robust performance in forecasting landslide displacement. Notably, the models' prediction accuracies are highest at moderate monitoring frequencies, surpassing those of daily and monthly monitoring frequencies. As monitoring frequency increases, the daily mean average error (MAE) experiences a rapid decline before stabilizing. Similar research results were also observed when analyzing the Bazimen landslide, corroborating that displacement prediction at moderate monitoring frequencies (approximately 7 to 15 days) yields superior accuracy compared to daily and monthly monitoring frequencies.